He invented deep learning — now he's killing LLMs

The man who co-invented deep learning (the mathematical foundation powering every modern AI system — from phone face-unlock to cancer screening) just raised $1.03 billion and declared that ChatGPT, Claude, and Gemini are building AI the fundamentally wrong way.

Yann LeCun — Turing Award winner (the Nobel Prize of computer science), co-inventor of CNNs (convolutional neural networks — the core algorithm behind every image-recognition app on your phone), and 12-year director of Meta's AI research lab — launched AMI Labs in early 2026. On March 10, 2026 the company announced a $1.03B (€890M) seed round at a $3.5 billion pre-money valuation — the largest seed round in European history. Backers include Jeff Bezos, NVIDIA, Samsung, Toyota, and Eric Schmidt.

His thesis, stated plainly: language models will never reach human-level intelligence because they learn only from text. Real intelligence starts in the physical world — by observing cause and effect, planning actions, and building internal simulations of how reality operates.

Why LeCun Says Today's AI Has a Fundamental Ceiling

Autoregressive language models (AI systems that generate text by predicting the next word, one step at a time — like ChatGPT or Gemini composing a sentence) are trained on text. But text is a thin slice of human knowledge: most of what humans know about the physical world was never written down.

A child learns that a glass falls and breaks by dropping it — not by reading about gravity. They understand fire burns by feeling heat, not by memorizing safety warnings. LeCun argues that without this embodied, cause-and-effect experience, language models will always hallucinate physical reasoning, fail at multi-step planning, and hit hard limits — no matter how many parameters they scale to.

The data backs him up. Meta's own V-JEPA 2 — a 1.2-billion-parameter world model prototype — currently scores just 44.5% on MVPBench (a test measuring whether AI can predict what physically happens next in a video). Humans score 85–95% on the same benchmark. The gap is massive. LeCun believes closing it requires a different architecture from the ground up — not simply scaling text data.

World Models vs. Language Models: The Core Difference

AMI Labs' core technology is called JEPA (Joint Embedding Predictive Architecture) — a method that predicts what happens next in compressed, abstract "concept space" rather than pixel-by-pixel. Think of it as AI learning to ask "roughly what will the outcome be?" rather than rendering every photon of the result. This filtering of unpredictable noise (the exact scatter pattern of broken glass is random and irrelevant for planning) is what LeCun believes makes world models far more sample-efficient and generalizable.

The $1.03B War Chest: Who Bet on LeCun

The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions (Jeff Bezos's personal investment vehicle). Strategic investors signal the breadth of applications:

• NVIDIA — the dominant AI chip manufacturer, betting world models will define the next hardware cycle
• Samsung — consumer electronics giant with deep interest in on-device, embodied AI
• Toyota Ventures — physical world understanding is prerequisite for autonomous vehicles and robotics
• Temasek — Singapore's sovereign wealth fund, providing state-level validation

Individual backers: Jeff Bezos, Mark Cuban, Eric Schmidt (former Google CEO), and Tim Berners-Lee (inventor of the World Wide Web). The founding team of 7 spans top AI institutions — CEO Alexandre LeBrun (serial AI entrepreneur), CSO Saining Xie (ex-Google DeepMind), and VP World Models Mike Rabbat. Offices are in Paris (HQ), New York, Montreal, and Singapore — a deliberately international structure positioning AMI as a European AI sovereignty play independent of US and Chinese cloud dominance.

First Products in 2027 — Here's What They Look Like

AMI Labs is not building a consumer chatbot. Its anchor domain partnership is with Nabla, a medical AI company — world models in healthcare can track patient conditions through time, model disease progression causally, and suggest interventions based on cause-and-effect reasoning rather than text pattern-matching on medical notes.

Other target verticals: industrial robotics (where physical cause-and-effect understanding is a safety requirement), industrial process control, and wearable devices. CEO LeBrun has stated explicitly: "About one year to get the first things we can use in the product." This is a multi-year research bet with real product milestones — not a vague open-ended research lab.

For context on the competitive landscape: World Labs (founded by Fei-Fei Li, the researcher who built ImageNet) raised $230M at a $1B valuation in 2024. Thinking Machines Lab (founded by former OpenAI CEO Mira Murati) reached a $12B valuation at seed. AMI Labs' $3.5B pre-money makes it the largest world-model startup in the world by capitalization.

LeCun departed Meta in November 2025. AMI Labs' website launched January 2026. The $1.03B closed by March 10, 2026. From departure to Europe's largest-ever seed close in under 5 months — that speed tells you how seriously the investment community takes his credibility.

Related Content — Get Started with Easy Claude Code | Free Learning Guides | More AI News